TY - JOUR
T1 - A comparative assessment of SARIMA, LSTM RNN and Fb Prophet models to forecast total and peak monthly energy demand for India
AU - Chaturvedi, Shobhit
AU - Rajasekar, Elangovan
AU - Natarajan, Sukumar
AU - McCullen, Nick
N1 - Funding Information:
The authors would like to acknowledge the funding received from the Department of Science and Technology , Government of India DST/TMD/UKBEE/2017/17 project: Zero Peak Energy Demand for India (ZED-I) and Engineering and Physics Research Council EPSRC project: EP/R008612/1 .
Funding Information:
Multiple Linear Regression (MLR), Support Vector Regression (SVR) and Artificial Neural Networks (ANN) are the most commonly used causal models for energy demand forecasting. MLR is the easiest to implement and provides estimates for parameter significance and model accuracy (Al-Hamadi and Soliman, 2005; Sharma et al., 2020; Li et al., 2017; Ur Rehman et al., 2017). However, as MLR approximates a linear function between the inputs and the output, it does not produce satisfactory results for datasets containing significant amounts of non-linearity and interaction effects (Ghalehkhondabi et al., 2017). In such situations, adaptive models like SVM and ANN are preferred (S. et al., 2012). ANN contains a collection of nodes that can mimic neurons' working in a biological brain (Hamza ç ebi et al., 2019). Several ANN architectures have been tested for energy demand forecasting as they are highly adaptive and efficient in learning complex dependencies between various model inputs and the output (Sahay et al., 2016; Azadeh et al., 2013; Ekonomou, 2010). For instance, Szoplik developed a multi-layer perceptron (MLP) ANN model to predict the natural gas consumption in Szczecin [Poland] considering the weather and temporal (month, day of the week, hour) effects. The authors tested several MLP configurations and identified MLP (22-36-1) containing 22 input neurons, 36 hidden neurons, and one output neuron as the most accurate model configuration (Szoplik, 2015). Similarly, Hamzaçebi et al. developed four adaptive ANN models to handle non-linear trends and seasonality effects in Turkey's energy demand data to develop monthly EDFs between 2015-2018 (Hamza ç ebi et al., 2019). Intelligent optimization techniques like grid search and nature-inspired heuristics have also been applied for optimal hyperparameter selection to enhance ANN performance (Khalid and Javaid, 2020; Jiang et al., 2020; Anand and Suganthi, 2017; Kankal and Uzlu, 2017). For instance, Muralitharan et al. compared the performance characteristics of a standalone ANN to a Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) assisted ANN such that GA and PSO selected optimal hyperparameters. Interestingly, both NN-GA and NN-PSO displayed higher prediction accuracy than conventional ANN. NN-GA faired better for short term load forecasting (hourly, daily), whereas NN-PSO performed better for long term EDS applications (months, years) (Muralitharan et al., 2018).The authors would like to acknowledge the funding received from the Department of Science and Technology, Government of India DST/TMD/UKBEE/2017/17 project: Zero Peak Energy Demand for India (ZED-I) and Engineering and Physics Research Council EPSRC project: EP/R008612/1.
Publisher Copyright:
© 2022 Elsevier Ltd
PY - 2022/9/30
Y1 - 2022/9/30
N2 - Selecting a suitable energy demand forecasting method is challenging due to the complex interplay of long-term trends, short-term seasonalities, and uncertainties. This paper compares four time-series models performance to predict total and peak monthly energy demand in India. Indian's Central Energy Authority's (CEA) existing trend-based model is used as a baseline against (i) Seasonal Auto-Regressive Integrated Moving Average (SARIMA), (ii) Long Short Term Memory Recurrent Neural Network (LSTM RNN) and (iii) Facebook (Fb) Prophet models. Using 108 months of training data to predict 24 months of unseen data, the CEA model performs well in predicting monthly total energy demand with low root-mean square error (RMSE 4.23 GWh) and mean absolute percentage error (MAPE, 3.4%), but significantly under predicts monthly peak energy demand (RMSE 13.31 GW, MAPE 7.2%). In contrast, Fb Prophet performs well for monthly total (RMSE 4.23 GWh, MAPE 3.3%) and peak demand (RMSE 6.51 GW, MAPE 3.01%). SARIMA and LSTM RNN have higher prediction errors than CEA and Fb Prophet. Thus, Fb Prophet is selected to develop future energy forecasts from 2019 to 2024, suggesting that India's annual total and peak energy demand will likely increase at an annual growth rate of 3.9% and 4.5%, respectively.
AB - Selecting a suitable energy demand forecasting method is challenging due to the complex interplay of long-term trends, short-term seasonalities, and uncertainties. This paper compares four time-series models performance to predict total and peak monthly energy demand in India. Indian's Central Energy Authority's (CEA) existing trend-based model is used as a baseline against (i) Seasonal Auto-Regressive Integrated Moving Average (SARIMA), (ii) Long Short Term Memory Recurrent Neural Network (LSTM RNN) and (iii) Facebook (Fb) Prophet models. Using 108 months of training data to predict 24 months of unseen data, the CEA model performs well in predicting monthly total energy demand with low root-mean square error (RMSE 4.23 GWh) and mean absolute percentage error (MAPE, 3.4%), but significantly under predicts monthly peak energy demand (RMSE 13.31 GW, MAPE 7.2%). In contrast, Fb Prophet performs well for monthly total (RMSE 4.23 GWh, MAPE 3.3%) and peak demand (RMSE 6.51 GW, MAPE 3.01%). SARIMA and LSTM RNN have higher prediction errors than CEA and Fb Prophet. Thus, Fb Prophet is selected to develop future energy forecasts from 2019 to 2024, suggesting that India's annual total and peak energy demand will likely increase at an annual growth rate of 3.9% and 4.5%, respectively.
KW - Energy demand forecasting
KW - Fb prophet
KW - LSTM RNN
KW - SARIMA
UR - http://www.scopus.com/inward/record.url?scp=85132534482&partnerID=8YFLogxK
U2 - 10.1016/j.enpol.2022.113097
DO - 10.1016/j.enpol.2022.113097
M3 - Article
AN - SCOPUS:85132534482
VL - 168
JO - Energy Policy
JF - Energy Policy
SN - 0301-4215
M1 - 113097
ER -